The Practice Of Statistics In The Life Sciences

Why Most Biology Students Mess Up Their Data Analysis (And How to Stop)

Statistical thinking in biology is harder than the math itself. You can run a t-test in your sleep, but picking the right test for a messy real-world dataset is where things fall apart. I have spent years watching people generate p-values that mean absolutely nothing because they never stopped to consider what their data actually looks like before feeding it into R. The core issue is that life science data violates every assumption classical statistics textbooks pretend exists. Your samples are never independent. Your variances are never equal. Your distributions are almost never normal, especially when you are dealing with things like gene expression counts, microbial abundances, or behavioral observations. The Practice Of Statistics In The Life Sciences isn't about memorizing formulas. It is about understanding what each test assumes and whether your data meets those assumptions.

Getting Started With The Practice Of Statistics In The Life Sciences

Start with visualization. Always. Before you run a single statistical test, plot your data. A boxplot, a strip chart, or a jitter plot will tell you more about what is going on than any normality test ever will. Shapiro-Wilk tests are notoriously overpowered with large sample sizes and underpowered with small ones. You will reject normality for trivial deviations when n equals 500, and fail to reject it when your data is clearly bimodal and n equals eight. Trust your eyes more than the test output. Once you have a sense of your data structure, pick your analysis based on three factors: your response variable type, your experimental design, and your sample size. These are not optional considerations. I watched a graduate student apply a standard linear model to count data last year. The residuals looked like garbage. Poisson or negative binomial regression would have been appropriate. She spent three months redoing the analysis after her advisor caught it during a routine check. For general linear models in R, start with the base functions and check assumptions using plot(model) before reaching for diagnostic packages. The residual versus fitted plot catches non-linearity. The Q-Q plot catches distributional violations. The scale-location plot catches heteroscedasticity. These four plots are the entire diagnostic toolkit for most standard models. Anything more elaborate is usually noise.

When your data has repeated measures or clustered structure, which it almost certainly does if you are working with anything biological, mixed effects models are the answer. The lme4 package handles this reasonably well for most standard designs. Specify your random effects based on the experimental structure, not based on what reduces your p-values. I once had someone remove a random effect because it made the model not converge, then reported results from a model that ignored the clustering in their data. That is how you get false positives. If convergence is a problem, try simplifying the random effects structure gradually or use glmmTMB for more complex variance structures. Multiple testing is another place where people consistently fail. Running fifty differential expression tests and looking at raw p-values is not science. It is fishing. Benjamini-Hochberg correction is the minimum standard for most life science applications. It controls the false discovery rate rather than the family-wise error rate, which is more appropriate when you are doing exploratory work with many comparisons. Holm-Bonferroni is more conservative and better for confirmatory analyses with fewer tests. There is no universal rule for which correction to use. It depends on whether you are generating hypotheses or testing them. Sample size planning is usually skipped entirely. You do not need a full power analysis with simulated data for every project, but estimating rough effect sizes from pilot data or published literature and doing a quick calculation saves you from publishing null results that were never detectable. G*Power handles standard designs. For complex designs, a quick simulation in R using your actual data structure gives a far more accurate estimate than any textbook formula.

One thing beginners rarely grasp is that statistical significance is not the same as biological significance. A drug might produce a statistically significant reduction in tumor volume with a large enough sample, but if the effect size is two percent, it is clinically irrelevant. Always report effect sizes alongside p-values. Cohen's d for t-tests, odds ratios for logistic regression, incidence rate ratios for count data. The number means something only when you know the magnitude. Missing data deserves more attention than it gets. Complete-case analysis is fine when data is missing completely at random, which is rare in biology. If samples are lost due to technical failure that correlates with the outcome, your results are biased. Multiple imputation using mice in R is the standard approach. It creates several complete datasets, analyzes each separately, and pools the results. It adds uncertainty properly rather than pretending the missing data never existed. I have seen papers get retracted because the authors ignored a systematic missingness pattern that their own data structure should have flagged. Finally, document everything. Your analysis pipeline should be reproducible from start to finish. Use R Markdown or Quarto to keep code, output, and narrative in one file. Version control your scripts even if you are working alone. The Practice Of Statistics In The Life Sciences improves dramatically when you treat your analysis as a living document rather than a sequence of clicks in a point-and-click interface. Most of the frustration people experience comes from losing track of which transformation they applied, which outliers they removed, and why they chose a particular model three weeks ago.

Resources like Clinical Biostatistics by Motulsky and Modern Applied Statistics with S by Venables and Ripley are useful, but the best learning happens when you work through real datasets. The ISwR package in R comes with biological datasets and exercises that mirror actual research problems. Working through those with a critical eye toward assumptions and alternatives is more valuable than reading another chapter on probability theory.